Microsoft’s Own Scientist Called AI Training Labor Theft

An unsealed 92-page brief in Manhattan federal court quotes Microsoft’s director of applied science calling AI training perhaps the largest theft of labor in human history. The document, filed by news publishers on Sept. 17, 2026, also cites Microsoft data that Copilot cut click-throughs to New York Times sites by 87% to 93% against ordinary Bing search.

Those lines sit in The New York Times Company v. Microsoft Corporation, case 1:23-cv-11195, now inside multidistrict litigation 25-md-3143 before Judge Sidney H. Stein. Microsoft and OpenAI still say training is fair use. The new public record is their own traffic study and their own product chiefs describing substitution.

The Memos Microsoft Tried to Keep Pink

The publishers, led by The New York Times and joined by the Daily News plaintiffs, Ziff Davis, the Center for Investigative Reporting, and The Intercept, lodged the public version of their 92-page combined summary judgment brief after redactions came off. The Times sued on Dec. 27, 2023. Cross-motions for summary judgment went in on Sept. 4, 2026; the quotes that had been sealed are what changed on Sept. 17.

Dr. Brent Hecht, Microsoft’s director of applied science, is the voice the brief puts on page one. Internal Microsoft documents, as quoted there, say millions of people would soon treat large models “hoovering up” their work as “an astonishing theft of unprecedented proportions,” and perhaps the largest theft of labor in human history. The same file adds that almost no one meant their work to be used that way, or is paid for it.

Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions.

Brent Hecht, Microsoft director of applied science, internal document quoted in the publishers’ brief

Hecht also wrote, the brief says, that a win for the companies on fair use would arguably “make a complete mockery of the idea of ‘fair use.’” A separate Hecht document, written soon after the lawsuit was filed, treats the products as useful because they copy patterns in the training text and, without a way to send money back down that chain, threaten the people who made it.

Jason Kint, chief executive of Digital Content Next, posted the unsealed pages the day they landed and said the companies’ own staff had written the Times’s lede.

woah. Court just unsealed Summary Judgment motions in New York Times v OpenAI / Microsoft.
Easy to see why the AI companies had so many redactions (pink). Their own people wrote the lede for NYT, the "largest theft of labor in human history"… "more and more substitutive…" /1 pic.twitter.com/gTRBiVuBOx

— Jason Kint (@jason_kint) September 17, 2026

That is the point the rest of the brief tries to lock in. Fair use in commercial copying turns, among other tests, on whether the new use replaces the original in the market. Book authors have already lost that fight in other courts when they could not show harm. This filing is built as the missing exhibit: the defendants describing replacement in their own words, then measuring it in Bing.

Copilot Cut Times Click-Throughs by 87% to 93%

A Microsoft document quoted in the brief names the spiral a “doom loop” that would hurt both the models and the web they feed on. “It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its ‘content supply chain,’” the document says. Another internal line is shorter: “LLMs are a product that destroys its supply chain.”

The publishers then point to Microsoft’s own comparison of Copilot’s “answer engine” with classic Bing search. The brief’s detailed statement of facts, SF1536, records these drops in click-through rate:

THE CLICK-THROUGH DROP

Copilot’s home page, the brief notes, greets users by saying that instead of clicking through links, they can talk through the question. An OpenAI engineer put the same idea more bluntly: “no matter how prominently we show the links, users won’t click.” The brief also says at least two million Copilot conversations copied and grounded outputs on the plaintiffs’ sites through retrieval-augmented generation, the live-lookup layer that sits on top of the trained model.

Satya Nadella, Microsoft’s chief executive, agreed under oath that chatting with bots “has substituted … giving you the information right there on the website on the AI platform versus needing to go to the underlying source.” He also said, in a deposition described in the same papers, that paywalled material should be licensed for grounding or training, and that he would have used Microsoft’s contract rights to force a retrain if he had known OpenAI had scraped behind a paywall.

What Nick Turley Wrote About Substitution

Nick Turley, who led ChatGPT, is the OpenAI voice the brief pairs with Hecht. In June 2023 he wrote that publishers faced an “existential threat.” Later notes, including February 2024, call the products “largely substitutive, period” and say they “will get more and more substitutive as they get better.”

Greg Brockman, OpenAI’s president, is quoted as calling the models “excellent at news,” “very good at any news task,” and especially good at finishing New York Times sentences. The brief’s substitution file is a stack of those lines, not a single leak.

WHAT THE BRIEF ATTRIBUTES TO DEFENDANTS

• Hecht, Microsoft: Training as astonishing theft, and a fair-use win as a mockery of the doctrine.
• Turley, OpenAI: An existential threat to publishers, with products that are substitutive and getting more so.
• Brockman, OpenAI: Models that are excellent at news and that complete Times sentences on point.
• Nadella, Microsoft: Chatbot answers have already stood in for visits to the source site.
• OpenAI research: “We train our networks to memorize the training data, that’s their objective.”

Hecht, asked about retrieval, agreed with a plain description: “In the simplest terms: AI search engines scrape answers from websites.” The publishers argue that this purpose, commercial and substitutive, knocks out the first fair-use factor, and that the click data knocks out the fourth. They cite the Supreme Court’s 2023 Warhol decision for the line that substitution is “copyright’s bête noire.”

That is not a finding by Judge Stein. It is the plaintiffs’ read of discovery, much of it still sitting in sealed exhibits. Microsoft’s reply is that Hecht does not speak for the company, and that Copilot is not a stand-in for journalism. OpenAI’s older public line is that regurgitation is a rare bug and that training on the public web is fair use.

Five Copying Stages, Including a Paywall Workaround

The brief asks the court to treat infringement as five separate acts, not one training run: acquiring copies, training on them, grounding outputs on them, reproducing them in answers, and swapping data between Microsoft and OpenAI under deals the publishers call horse trading. Project Taxi and Project Mango are the names it gives that swap. OpenAI, it says, sent Microsoft the full GPT-3 training set; Microsoft sent news text back.

THE FIVE COPYING STAGES

• Acquisition: Scrapes, Common Crawl dumps, and licensed archives copied without the publishers’ leave.
• Training: Those copies used in pre-training and mid-training, including a mid-training pool with more than 91,692 copies of the plaintiffs’ works.
• Grounding: Live lookups that pull site text into Copilot and ChatGPT answers.
• Output copying: Verbatim and near-verbatim stretches of articles in chat logs.
• Horse trading: Dataset exchanges, including Project Mango’s at least 160,903 unique plaintiff works.

Scale, as the brief counts it, is not a vibe. A post-filtered Common Crawl set used in training listed 2,064,805 articles from nytimes.com and 2,182,079 from chicagotribune.com. WebText2, an earlier OpenAI scrape, held 6,552 Times works, 18,609 Daily News works, and 66,780 Ziff Davis works. OpenAI also took the New York Times Annotated Corpus from the Linguistic Data Consortium, “nearly every article published in The New York Times between January 1, 1987 and June 19, 2007,” more than 1.8 million pieces.

WHAT THE BRIEF COUNTS

Acquisition is the chapter that will travel. When researcher Nick Ryder told Brockman about “a hack to get around nytimes paywall” to help scrape the site, Brockman replied “ah nice.” OpenAI’s corporate witness on the topic, the brief says, did not know of any method then in use for spotting paywalled text in the datasets. Researchers also stripped copyright management information so models would not spit copyright notices back at users. Hecht called a related filter an “accidental cover up,” because people with rights in the work would see less of what was used to train.

OpenAI has said it offers publishers a crawler opt-out and that The Times used it in August 2023. The brief’s claim is that the copies at issue were already in the vaults, and that Common Crawl and anonymous scrapes went around robots.txt anyway. OpenAI’s public fair-use and opt-out position still treats training on public web text as lawful and regurgitation as a bug to drive toward zero.

Microsoft Says Hecht Was Paid to Disagree

Microsoft’s first move on Hecht was not to argue with the sentences. It was to shrink the speaker. A Microsoft spokesperson said the comments “reflect one employee’s individual perspective, are not a legal analysis, and do not represent the company’s views.” In a separate filing, Microsoft told the court Hecht is a research academic who also worked at Northwestern University, was not a decision-maker, and was employed “to present divergent and asymmetric perspectives.”

That label is a real fact about his job. It does not reach Nadella’s deposition, Turley’s notes, Brockman’s “ah nice,” or the Copilot click study, which are not Hecht’s thought experiments. Microsoft’s legal papers still analogize training to a student reading, and still say Copilot is not a substitute for the publishers’ journalism. Spokesperson Alex Haurek said the company’s position is the one in those filings.

Steven Lieberman, counsel for the New York Daily News, said the unsealed evidence shows OpenAI and Microsoft “knew that what they were doing was wrong.” OpenAI did not return requests for comment on the Sept. 17 brief. The live commercial argument, once the theft metaphor is stripped off, is whether retrieval and answer engines have to pay for the pages they read out loud. Kint later put it as a licensing fight, especially for RAG, not a seminar on whether weights can learn from text.

On Sept. 1 the Justice Department Backed Training

Sixteen days before the quotes became public, the United States filed a statement of interest on fair use in the same MDL. Associate Attorney General Stanley E. Woodward Jr. and Assistant Attorney General Brett Shumate signed for the department. The government said the “training of AI models on copyrighted material,” in and of itself, “does not violate copyright laws.”

Rules of law that make it significantly more difficult to develop a robust AI industry in the United States therefore threaten national security and give a competitive advantage to foreign adversaries who are not so encumbered.

Statement of Interest of the United States, Sept. 1, 2026

The filing leans on executive orders from Jan. 23, 2025, and June 2, 2026, on American AI leadership. It also draws a line the publishers will test. American creators, it says, should be protected from infringing outputs without blocking lawful innovation, and training should be analyzed apart from what a chatbot later emits. In that split, a memorized Times paragraph in a chat log is a different question from the gradient updates that built the model.

The department warned that a licensing duty could leave only the largest firms able to pay, and would function as a subsidy for legacy outlets. Publishers read the same brief as Washington taking the side of companies that already copied the library. The SOI is not a ruling. It is the first time the Justice Department has told a court, in these cases, that training on the written record should be treated as fair use.

Bartz’s $1.5 Billion Did Not Decide News Scraping

Two Northern District of California decisions from June 2025 are the fair-use floor OpenAI now cites. On June 23, 2025, in Bartz v. Anthropic, the court treated using lawfully obtained books to train a model as fair use on that record, and treated a digital library of pirated files as a separate, unprotected act. On July 20, 2026, the same court approved a $1.5 billion class settlement covering about 482,000 works, described there as the largest copyright class settlement on record, for the piracy side rather than for training itself.

On June 25, 2025, in Kadrey v. Meta Platforms, Meta won summary judgment on the training theory because the named authors had not put up proof on the fourth factor, market harm. The court said Meta did not win a holding that copying books to train is always fair use. It won because the plaintiffs had not shown the market fact. That sentence is why the Times brief opens with Hecht and closes with Bing click graphs.

THE CASE CLOCK

1. December 27, 2023: The New York Times sues Microsoft and OpenAI in the Southern District of New York.
2. June 23, 2025: Bartz v. Anthropic treats lawful training as fair use and keeps the pirated-library claims alive.
3. June 25, 2025: Kadrey v. Meta grants the company summary judgment on training after a weak market-harm showing.
4. June 25, 2026: The Times amends, dropping contributory infringement and trademark dilution after the Supreme Court’s Cox decision.
5. July 20, 2026: Bartz’s $1.5 billion piracy settlement wins final approval.
6. September 1, 2026: The Justice Department files its statement of interest backing training as fair use.
7. September 4, 2026: The parties file cross-motions for summary judgment.
8. September 17, 2026: The publishers file the public, less-redacted brief that quotes Hecht, Turley, Brockman, and the Copilot click study.

An independent docket tracker listed 140 U.S. AI lawsuits as of Sept. 20, 2026, with OpenAI named in 34 of them. Those other complaints can now point to a Manhattan record in which the model makers’ own applied-science lead used the word theft, their ChatGPT chief used the word substitutive, and their search partner measured a collapse in clicks. Judge Stein has not decided whether any of that defeats fair use. The underlying memos and datasets remain, in large part, under seal.

What is public is enough to change the shape of the next argument. Training-as-reading still has a government brief and two California training rulings behind it. Replacement-as-product now has Microsoft’s click tables and OpenAI’s product notes on the other side of the same docket. The exhibits Hecht actually typed are still in the vault. The sentences the publishers were allowed to print are already in every later complaint that has to prove a market.

Disclaimer: This article is news reporting and analysis of public court filings, company statements, and related case law. It is for information only and is not legal advice, does not assess the merits of any claim or defense, and should not be relied on as counsel about copyright, fair use, licensing, or litigation strategy. Readers with a legal question about training data, outputs, or a live dispute should consult a qualified copyright attorney admitted in the relevant jurisdiction before acting. Figures, quotes, and case statuses are those in the cited filings and statements as of the dates given in the piece and may change as the court rules or as the parties file further papers.